tests: avoid collecting consistency tests that are guaranteed to skip in CI

[njzjz-bot]

Summary

Several cross-backend consistency suites collect large numbers of test items that are guaranteed to skip in a given CI job (for example, backend- or device-specific tests collected in jobs where the backend/device is not meant to be exercised).

This inflates collected counts, increases scheduling overhead, and makes CI logs noisy without improving coverage.

Motivation

Recent CUDA CI analysis showed a recurring pattern in heavy consistency suites:

• large numbers of collected items
• a substantial fraction ending up as skipped
• backend/device combinations that are clearly not relevant for the current job

This looks like a structural test-selection problem rather than an isolated slow test.

Proposed direction

Avoid collecting tests that are known in advance to be unsupported or irrelevant for the current CI job.

Possible approaches include:

• backend-specific markers and CI-side test selection
• backend-specific test classes/modules instead of broad classes with runtime skips
• collection-time filtering rather than skipTest() / runtime skip where feasible
• clearer separation between CPU-only, CUDA-relevant, and backend-specific suites

Scope

This should be applied broadly across heavy consistency suites rather than focusing on a single file.

Acceptance criteria

• CUDA CI no longer collects large numbers of backend/device tests that are guaranteed to skip
• collected and skipped counts decrease significantly in heavy consistency jobs
• CI coverage remains intentional and documented

---

Authored by OpenClaw (model: gpt-5.4)